ORION-CMR Delivers On-Scanner Cardiac MRI Reports in 90 Seconds

Cardiac magnetic resonance imaging can characterize ventricular function, myocardial scar, and several forms of structural or ischemic disease, but its clinical use is constrained by a multi-step workflow. Acquisition is followed by sequence identification, image post-processing, measurements, interpretation, and report writing. Most prior AI systems target one of those steps. ORION-CMR instead makes the integration itself the contribution: a scanner-native system intended to turn a completed CMR examination into structured analysis and a generated clinical impression.

Core Contribution

The authors present ORION-CMR, short for On-scanner Reporting with Integrated fOunda-tioN Model, as an end-to-end CMR foundation model. It was pretrained on 12,896,733 CMR images drawn from 9,258 studies, then used for sequence classification, ventricular-function assessment, late gadolinium enhancement (LGE) detection, binary and multiclass disease classification, and report generation through a local large language model.

That breadth is the meaningful distinction from task-specific pipelines. The paper's claim is not merely that one classifier improves a benchmark, but that a shared image representation can supply the intermediate findings needed to produce an examination-level impression. Figure 1 depicts this scanner-native workflow, from CMR input through analysis to an automated report.

Technical Approach

The available paper material describes a framework that operates on the scanner and generates its report locally rather than requiring a separate remote reporting service. The system couples its imaging outputs to a local large language model for impression generation. This design matters in a clinical setting because report generation must reflect more than a single disease label: the model needs to connect sequence-level identification, ventricular measurements, and LGE-related findings into a coherent interpretation.

The paper evaluates both binary normal-versus-abnormal classification and multiclass disease classification, with a cohort containing normal examinations, congenital heart disease, dilated cardiomyopathy, and myocardial infarction. That case mix is clinically more informative than a single-diagnosis test, although it remains a limited slice of routine CMR practice.

The report evaluation is framed as expert-reader agreement rather than as a text-similarity score. Figure 2 pairs original and generated impressions and shows the distribution of review scores across 68 cases. Figure 3 further separates concordant, intermediate, and discordant examples, which is the right qualitative lens for a reporting system: a plausible-sounding report is insufficient if its clinical conclusion diverges from reader assessment.

Results and Analysis

The authors report approximately 90-second end-to-end analysis and reporting. In the clinical cohort, ORION-CMR achieved an area under the ROC curve of 0.96 for normal-versus-abnormal classification and 0.88 for multiclass disease classification. Generated reports agreed with expert interpretation in 81.4% of cases. The paper also states that the system exceeded supervised baselines and the prior CMR-FM foundation model on its evaluated tasks, including LGE classification and scar segmentation, but the available material does not provide the corresponding benchmark values or margins.

The 90-second workflow is the practical result to watch. A model that only improves an offline segmentation score has a narrow operational role; combining several outputs into an on-scanner impression could shorten the path from imaging to review. But the reported 81.4% agreement also means the generated impression cannot be treated as autonomous reporting. The discordant examples shown in the paper reinforce that expert oversight remains necessary.

Evidence and Boundaries

The evidence supports feasibility in a real clinical cohort, not broad deployment. The evaluation covers 68 subjects and four diagnostic groups, while CMR referrals include a wider range of protocols, pathologies, image-quality failures, and institutional reporting conventions. The paper reports strong discrimination metrics and a clinically relevant reader assessment, but the available results do not establish performance across additional sites, prospective scanner deployment, or downstream effects on reporting time and patient care. ORION-CMR is therefore best read as an integrated decision-support workflow with encouraging clinical validation, rather than proof that automated reports can replace CMR experts.